MacWilliams identities for terminated convolutional codes Please share

advertisement
MacWilliams identities for terminated convolutional codes
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Forney, G. David. “MacWilliams Identities for Terminated
Convolutional Codes.” 2010 IEEE International Symposium on
Information Theory Proceedings (ISIT). 1105–1109. © Copyright
2010 IEEE
As Published
http://dx.doi.org/10.1109/ISIT.2010.5513699
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Version
Final published version
Accessed
Fri May 27 00:19:14 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/73077
Terms of Use
Article is made available in accordance with the publisher's policy
and may be subject to US copyright law. Please refer to the
publisher's site for terms of use.
Detailed Terms
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
MacWilliams Identities for
Terminated Convolutional Codes
G. David Forney, Jr.
Laboratory for Information and Decision Systems
Massachusetts Institute of Technology
Cambridge, MA 02139
Email: forneyd@comcast.net
Abstract—Shearer and McEliece [8] showed that there is no
MacWilliams identity for the free distance spectra of orthogonal
linear convolutional codes. We show that on the other hand
there does exist a MacWilliams identity between the generating
functions of the weight distributions per unit time of a linear
convolutional code C and its orthogonal code C ⊥ , and that
this distribution is as useful as the free distance spectrum for
estimating code performance. These observations are similar to
those made recently by Bocharova et al. [1]; however, we focus
on terminating by tail-biting rather than by truncation.
I. I NTRODUCTION
Finding a MacWilliams-type identity for convolutional
codes is a problem of long standing [8]. For a linear timeinvariant convolutional code C over a finite field, the most
commonly studied distance distribution is the free (Hamming)
distance spectrum, namely, the distribution of (Hamming)
weights of codewords in C that start and end in the zero state
without passing through an intermediate zero state. Shearer
and McEliece [8] showed by example that the free distance
spectrum of C does not in general determine that of C ⊥ , and
therefore that there could be no MacWilliams identity for such
distributions.
Gluesing-Luerssen and Schneider (GLS) have recently formulated [6] and proved [7] a MacWilliams-type identity for
convolutional codes involving the Hamming weight adjacency
matrix (HWAM) of a convolutional code C and the HWAM of
its orthogonal code C ⊥ . In [3], [4], the GLS result was proved
in a different way, and generalized to various kinds of weight
adjacency matrices and to group codes defined on graphs.
More recently, Bocharova, Hug, Johannesson and
Kudryashov [1] have proved a different MacWilliamstype identity for truncations of a convolutional code C and
its orthogonal code C ⊥ , and have shown that by letting
the truncation length become large, an approximation to
the free distance spectrum can be obtained. In this paper,
which is mostly based on [4], we derive similar results for
weight distributions of block codes obtained by various kinds
of termination procedures, of which we regard tail-biting
as the nicest. We argue that these alternative distributions
are just as useful for estimating code performance as the
free distance spectrum. These results effectively answer the
original question posed by Shearer and McEliece [8].
978-1-4244-7892-7/10/$26.00 ©2010 IEEE
II. T ERMINATED CONVOLUTIONAL CODES
A general method for approximating the free distance
spectrum of a linear convolutional code C is to derive a
series of block codes CN of length N from C by some
sort of termination procedure, and then to study the distance
distributions of CN as N → ∞. As we shall see in Section
III, for any of the termination methods below, the distance
distribution of CN , normalized by N , approaches the free
distance spectrum of C for dfree ≤ d < 2dfree . However,
the usual termination methods are problematic if we are
also interested in the distance distribution of the orthogonal
convolutional code C ⊥ .
For example, the most common termination method is to
take the subcode C[0,N ) of C, consisting of all sequences in
C whose support is contained in the interval [0, N ) = {k ∈
Z | 0 ≤ k < N } (i.e., all code sequences that pass through
the zero state at times 0 and N , restricted to [0, N )), which
is effectively a block code of length N time units. As a
subcode of C, C[0,N ) has at least the minimum (free) distance
of C. However, the orthogonal code to C[0,N ) is the projection
(C ⊥ )|[0,N ) of the orthogonal convolutional code C ⊥ onto the
interval [0, N ) (i.e., all orthogonal code sequences that pass
through any state at times 0 and N , restricted to [0, N )). In
general, a projection (C ⊥ )|[0,N ) has low-weight codewords, no
matter how large N becomes.
Bocharova et al. [1] have considered another kind of terminated code that they call a truncated code, which we will
denote by C[0,N ) . Such a code may be described as the
subcode C[0,∞) projected onto [0, N ): i.e.,
C[0,N ) = (C[0,∞) )|[0,N ) .
By projection/subcode duality, the dual code (C[0,N ) )⊥ is
((C[0,∞) )|[0,N ) )⊥ = ((C[0,∞)) )⊥ )[0,N ) = ((C ⊥ )|[0,∞)) )[0,N ) ;
i.e., the subcode defined on [0, N ) of the projection of the
orthogonal convolutional code C ⊥ onto [0, ∞). Alternatively,
(C[0,N ) )⊥ may be defined as a reverse-truncated code [1]
(C ⊥ )[0,N ) = ((C ⊥ )(−∞,N )) )|[0,N ) .
Thus there is a MacWilliams identity between the weight
distributions of C[0,N ) and (C ⊥ )[0,N ) . Moreover, unlike the
1105
ISIT 2010
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
subcode C[0,N ) or the projection C|[0,N ) , a truncated code
has the same rate as C. However, truncated codes have small
minimum distance.
A more elegant method of terminating a convolutional code
C is via tail-biting. The tail-biting terminated code C||[0,N ) of
the convolutional code C on the interval [0, N ) is the set of
all codewords in C that pass through the same state at times
0 and N , restricted to [0, N ). For large enough N , the tailbiting code C||[0,N ) has the same minimum distance as C;
moreover, C||[0,N ) has the same rate as C. Most importantly,
the orthogonal code to C||[0,N ) is (C ⊥ )||[0,N ) , the tail-biting
terminated code of C ⊥ on the same interval [2]. Thus there
is a MacWilliams identity between the weight distributions
of C||[0,N ) and (C ⊥ )||[0,N ) . Finally, we will see that these
distributions approach the free distance spectra of C and C ⊥
nicely as N → ∞.
The minimum distance of this block code is 2, less than that
of C ⊥ , although its rate is higher.
The truncated code C[0,4) is the (8, 4) binary linear block
code generated by
Example 1 (rate-1/2 4-state binary linear convolutional code).
Consider the rate-1/2 binary linear time-invariant convolutional
code C with degree-2 generators (1 + D2 , 1 + D + D2 ), in
standard D-transform notation. A minimal encoder for this
code is the linear time-invariant system with impulse response
(11, 01, 11, 00, . . .), which has the 4-state trellis section shown
in Figure 1(a). The orthogonal convolutional code C ⊥ is
the rate-1/2 binary linear convolutional code with generators
(1 + D + D2 , 1 + D2 ), which has a minimal 4-state linear
encoder with impulse response (11, 10, 11, 00, . . .), and the 4state trellis section shown in Figure 1(b).
which has the same parameters.
The tail-biting terminated code C||[0,4) is the (8, 4) binary
linear block code generated by
00H 00
00
HH
11
11
10H 00H
10
@H
H
01@ 01H
01
@
10
10
11 01 @11
00H 00
00
HH
11
11
01H 00H
01
@H
H
10@ 10H
10
@
01
01
11 10 @11
(a)
(b)
Fig. 1. Trellis sections of (a) rate-1/2 4-state binary convolutional code C;
(b) orthogonal code C ⊥ .
We now consider various methods of terminating this convolutional code C with a block length of N = 4. The subcode
C[0,4) is the (8, 2) binary linear block code generated by the
two generators
11 01 11 00
00 11 01 11
The minimum distance of this block code is the same as that
of C, namely 5, although its rate is lower. The orthogonal
code to the subcode C[0,4) is the projection (C ⊥ )|[0,4) of the
orthogonal convolutional code C ⊥ , which is the (8, 6) binary
linear block code generated by the six generators
11
10
11
00
00
00
00
11
10
11
00
00
00
00
11
10
11
00
00
00
00
11
10
11
11
00
00
00
01
11
00
00
11
01
11
00
00
11
01
11
The minimum distance of this block code is 2, but its rate is
the same as that of C. Its orthogonal code (C ⊥ )[0,4) is the
(8, 4) binary linear block code generated by
11
10
11
00
11
00
11
01
00
11
10
11
01
11
00
11
00
00
11
10
11
01
11
00
00
00
00
11
00
11
01
11
whereas the orthogonal tail-biting terminated code (C ⊥ )||[0,N )
is the (8, 4) binary linear block code generated by the four
generators
11 10 11 00
00 11 10 11
11 00 11 10
10 11 00 11
Both of these codes have a minimum distance of only 2 (e.g.,
for paths such as 01 00 01 00 from state 10 to state 10).
However, for N ≥ 10, it turns out that the minimum distance
of both tail-biting terminated codes is 5, the same as the
minimum distance of C or C ⊥ .
III. F REE DISTANCE SPECTRA FOR CONVOLUTIONAL
CODES FROM TERMINATED CODES
Let us now consider how the free distance spectrum of a
linear time-invariant convolutional code C may be derived from
the weight distribution of a terminated code of length N as
N → ∞.
Without loss of generality, we may assume that C is
generated by a minimal encoder, which is necessarily noncatastrophic: i.e., the unique state sequence associated with
the all-zero code sequence is the all-zero state sequence.
Consequently, the lowest-weight words of a terminated code
as N → ∞ must be those that pass through the zero state
almost all of the time. These code sequences are as follows,
for the various termination methods we have considered:
• If we terminate to the subcode C[0,N ) , then code sequences start and end in the zero state, and the lowestweight sequences correspond to the lowest-weight sequences in the free distance spectrum. If the minimum
1106
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
•
•
•
free distance is dfree , then for dfree ≤ d < 2dfree
there will be approximately N × Nd sequences in the
terminated code of weight d, where Nd is the number
of code sequences of weight d in the free distance
spectrum of C. Thus, for dfree ≤ d < 2dfree , the weight
distribution per unit time of C is the limit of the weight
distribution of C[0,N ) normalized by (divided by) N as
N → ∞. For d ≥ 2dfree , there will be overcounting—
e.g., two sequences of weight dfree may be counted as
one of weight 2dfree — but we will argue below that
such overcounting should not affect estimates of code
performance.
If we terminate to the projection C|[0,N ) , then code
sequences can start and end in any state, and there will
be low-weight sequences starting with a low-weight state
transition s → 0, remaining in state 0 for nearly N
time units, and then ending with a low-weight transition
0 → s , where s and s are not both 0. Thus the minimum
distance of C|[0,N ) will be less than dfree for all N . However, the number of such low-weight sequences remains
constant, so after normalization we will eventually see
the same normalized weight distribution as for C[0,N ) .
If we terminate to the truncated code C[0,N ) , then by
the same argument we will eventually see the same
normalized weight distribution. In this case the total
weight of a code sequence starting in the zero state,
remaining there for nearly N time units, and then ending
with a low-weight transition 0 → s, is only that of
the low-weight transition 0 → s. However, again the
number of such low-weight sequences remains constant,
so after normalization we will eventually see the correct
normalized weight distribution.
If we terminate to the tail-biting code C||[0,N ) , then by
the same argument we will eventually see the same
normalized weight distribution. Note however that in this
case the total weight of a code sequence starting with
a low-weight transition s → 0, remaining in the zero
state for nearly N time units, and then ending with a
low-weight transition 0 → s, must be at least dfree , since
the ending sequence (corresponding to the state transition
0 → s) followed by the starting sequence (corresponding
to s → 0) must be a code sequence. Thus the minimum
distance of C||[0,N ) must equal dfree for large enough N .
We conclude that as N → ∞ the normalized weight
distribution of any of these terminated codes approaches the
free distance spectrum of C for dfree ≤ d < 2dfree . However,
only the tail-biting termination has the same rate as C and, for
N large enough, the same minimum distance dfree .
Finally, we argue that the normalized weight distribution of
any of these terminated codes CN must yield the same estimate
of code performance over N time units as the free distance
spectrum of C, if these estimates are accurate. The probability
of error event P (E) of C per unit time may be estimated using
the free distance spectrum. The probability of any error in
N time units is then estimated as N P (E). If this is a good
estimate (implying N < 1/P (E)), then the probability of two
or more error events in N time units must be negligible. But
the probability of any error in decoding C over N time units
is essentially the same as the probability of block decoding
error in decoding CN , which may be estimated by the weight
distribution of N , which counts codewords that include two
or more error events. If the probability of two or more error
events in N time units is negligible, then an estimate based on
the weight distribution of CN must approximately agree with
an estimate based on the free distance spectrum of C.
IV. W EIGHT ADJACENCY MATRICES AND WEIGHT
GENERATING FUNCTIONS
We now show how Hamming weight generating functions
for terminations of a linear time-invariant convolutional code C
may be derived from the Hamming weight adjacency matrix of
a minimal linear time-invariant encoder for C. This will allow
us to state MacWilliams identities for terminated convolutional
codes, and to estimate code performance.
Given a linear time-invariant encoder for a convolutional
code C with state space S and symbol alphabet A, the
Hamming weight adjacency matrix (HWAM) is the matrix
Λ(x) indexed by S × S whose elements are
xw(a) ,
Λss (x) =
a∈T (s,s )
where x is an indeterminate, T (s, s ) is the subset of symbols a ∈ A such that the state/symbol transition (“branch”)
(s, a, s ) ∈ S × A × S actually occurs in the encoder, and
w(a) is the Hamming weight of the symbol a ∈ A.
Example 1 (cont.). For the rate-1/2 binary convolutional code
C of Example 1, the HWAM of the encoder of Figure 1(a) is
Λ(x)
=
s/s
00
10
01
11
00
1
0
x2
0
10
x2
0
1
0
01 11
0
0
x x
0
0
x x
For the orthogonal code C ⊥ , the HWAM of the encoder of
Figure 1(b) is
Λ̂(x)
=
s/s
00
10
01
11
00
1
x2
0
0
10 01
0 x2
0
1
x 0
x 0
11
0
0
x
x
which in this case is the transpose of Λ(x).
It is shown in [6], [7], [3], [4] how the HWAM Λ(x) of
a minimal encoder for C determines the HWAM Λ̂(X) of a
minimal encoder for C ⊥ and vice versa via a MacWilliamstype identity, but we will not need that result here.
Now it is easy to see that if we take N consecutive trellis
sections of a minimal encoder for C as a single section, then
the HWAM of this length-N trellis section is simply the N th
power ΛN (x) of the basic HWAM Λ(x).
1107
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
⎡
1 + 2x5 + x6
⎢
x3 + 2x4 + x5
Λ4 (x) = ⎢
⎣ x2 + x3 + x4 + x7
x3 + 2x4 + x5
Fig. 2.
x2 + x3 + x4 + x7
x2 + x3 + x5 + x6
x2 + x4 + 2x5
2
x + x3 + x5 + x6
x3 + 2x4 + x5
2x3 + x4 + x6
2
x + x3 + x5 + x6
2x3 + x4 + x6
⎤
x3 + 2x4 + x5
2x3 + x4 + x6 ⎥
⎥.
2
x + x3 + x5 + x6 ⎦
2x3 + x4 + x6
HWAM Λ4 (x) of a section of N = 4 time units of Example 1 code C.
Example 1 (cont.). Given the HWAM Λ(x) above for a
minimal encoder of our example code C, the HWAM of a
section consisting of N = 2 time units of this code is
⎡
⎤
1 x2 x3 x3
⎢ x3 x x2 x2 ⎥
⎥
Λ2 (x) = ⎢
⎣ x2 x4 x x ⎦ .
x3 x x2 x2
This shows that there is exactly one path from each state at
time k to each state at time k + 2, and that the minimum
Hamming weight of any of these paths (other than the zero
path) is 1.
For a section consisting of N = 4 time units of this code,
the HWAM Λ4 (x) is given in Fig. 2. This HWAM shows that
there are four paths from each state at time k to each state at
time k + 4, and that the minimum nonzero Hamming weight
of any of these paths is 2.
The weight generating functions of various terminated codes
of C can now be read from these weight adjacency matrices.
Since the subcode C[0,N ) is the set of all sequences in C that
pass through the zero states at times 0 and N , its weight
generating function is simply the (0, 0) element of ΛN (x).
Similarly, since the projection C|[0,N ) is the set of all sequences
in C that pass through any states at times 0 and N , its weight
generating function is the sum of all elements of ΛN (x).
Since the truncated code C[0,N ) is the set of all sequences
in C that pass through the zero state at times 0 and any state
at time N , its weight generating function is the sum of all
elements in the first row of ΛN (x). Similarly, the weight
generating function of C[0,N ) is the sum of all elements in
the first column of ΛN (x).
Finally, since the tail-biting termination C||[0,N ) is the set
of all sequences in C that pass through the same states in
S0 and SN , its weight generating function is the sum of
all diagonal elements of ΛN (x); i.e., its trace Tr (ΛN (x)).
Since C||[0,N ) and (C ⊥ )||[0,N ) are orthogonal block codes, there
is a MacWilliams identity between their weight generating
functions.
Example 1 (cont.). For the rate-1/2 binary convolutional code
of Example 1, the Hamming weight generating function of
the the tail-biting termination C||[0,4) of length 4 is the trace
of Λ4 (x), namely 1 + 2x2 + 4x3 + x4 + 4x5 + 4x6 . Since
Λ̂4 (x) is the transpose of Λ4 (x), the orthogonal tail-biting
terminated code (C ⊥ )||[0,4) is an equivalent code with the same
Hamming weight generating function. It is easy to check that
this Hamming weight generating function is indeed invariant
under the MacWilliams transform.
Using tail-biting terminated codes, and normalizing the
weight distribution by dividing by N , we have that the generating function of the normalized Hamming weight distribution
of C is
1
Tr (ΛN (x)).
gC (x) = lim
N →∞ N
Moreover, there is a MacWilliams identity between gC (x) and
gC ⊥ (x). The performance of C is determined by gC (x), and
that of C ⊥ by gC ⊥ (x). (Similar observations are made in [1],
using truncated codes.)
Example 1 (cont.). For a section consisting of N = 16 time
units of the rate-1/2 binary convolutional code C of Example
1, the HWAM Λ16 (x) (modulo x8 ) is given in Fig. 3 at the
top of the next page. Notice that
Tr (Λ16 (x)) = 1 + 16x5 + 32x6 + 64x7 + · · · ,
so that normalizing the distribution by dividing by N = 16
already gives the precise free distance spectrum of C for d < 8,
namely x5 + 2x6 + 4x7 + · · ·. Thus the convergence to the
limiting generating function gC (x) is rapid and exact. This
property of tail-biting terminations is not shared by other kinds
of terminations.
It appears that the behavior of gC (x) might be analyzed by
using an extension of Perron-Frobenius theory to generating
function matrices, as in [5]; however, we have not attempted
such an analysis.
Example 2 (cf. [8], [1]). The two codes proposed by Shearer
and McEliece [8] for their counterexample make an excellent
example. The first code is a rate-1/3 binary linear timeinvariant convolutional code C1 generated by the degree-1
generators (1, 1 + D, D), i.e., C1 is generated by a minimal
encoder with impulse response (110, 011, 000, . . .), whose
trellis section is shown in Figure 4(a). The HWAM of this
encoder is
1 x2
.
Λ1 (x) =
x2 x2
0 H 000 0
H
011 H
H
011
H
1
101 H 1
(a)
0 H 000 0
H
001 H
H
111
H
1
110 H 1
(b)
Fig. 4. Trellis sections of (a) rate-1/3 2-state binary convolutional code C1 ;
(b) similar code C2 .
The second code is a rate-1/3 binary linear time-invariant
convolutional code C2 generated by the degree-1 generators
(D, D, 1+D), i.e., C2 is generated by a minimal encoder with
1108
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
1 + 14x5 + 25x6 + 44x7
x3 + 2x4 + 4x5 + 8x6 + 16x7
x2 + x3 + 2x4 + 4x5 + 8x6 + 29x7
x3 + 2x4 + 4x5 + 8x6 + 16x7
Fig. 3.
x2 + x3 + 2x4 + 4x5 + 8x6 + 29x7
x5 + 3x6 + 8x7
x4 + 2x5 + 5x6 + 12x7
x5 + 3x6 + 8x7
x3 + 2x4 + 4x5 + 8x6 + 16x7
x6 + 4x7
x5 + 3x6 + 8x7
x6 + 4x7
x3 + 2x4 + 4x5 + 8x6 + 16x7
x6 + 4x7
x5 + 3x6 + 8x7
x6 + 4x7
HWAM Λ16 (x) (modulo x8 ) of a section of N = 16 time units of Example 1 code C.
impulse response (001, 111, 000, . . .), whose trellis section is
shown in Figure 4(b). The HWAM of this encoder is
1 x
.
Λ2 (x) =
x3 x2
Since the weights of the 0 → 0 and 1 → 1 transitions are the
same for C1 and C2 , and since the sums of the weights of the
0 → 1 and 1 → 0 transitions are the same, it is evident that the
weight distributions of the subcodes (C1 )[0,N ) and (C2 )[0,N )
are the same for all N , and that the free distance spectra of
C1 and C2 are also the same. For the same reason, the weight
distributions of the tail-biting terminated codes (C1 )||[0,N ) and
(C2 )||[0,N ) are the same for all N .
However, the weight distributions of the projections
(C1 )|[0,N ) and (C2 )|[0,N ) are not the same even for N = 1. It
follows that the weight distributions of the subcodes (C1⊥ )[0,N )
and (C2⊥ )[0,N ) of their orthogonal codes C1⊥ and C2⊥ are not
the same, and therefore that their free distance spectra are not
the same; this was the point of Shearer and McEliece [8].
On the other hand, since the weight distributions of the tailbiting terminated codes (C1 )||[0,N ) and (C2 )||[0,N ) are the same
for all N , it follows that the weight distributions of the tailbiting terminated codes (C1⊥ )||[0,N ) and (C2⊥ )||[0,N ) are the
same for all N .
Since the performance of C1⊥ and C2⊥ can be analyzed from
these weight distributions, it follows that the performance of
C1⊥ and C2⊥ is effectively the same, despite the difference in
their free distance spectra.1
R EFERENCES
[1] I. E. Bocharova, F. Hug, R. Johannesson and B. Kudryashov, “A note
on convolutional codes: Equivalences, MacWilliams identity, and more,”
submitted to IEEE Trans. Inform. Theory, 2009.
[2] G. D. Forney, Jr., “Codes on graphs: Normal realizations,” IEEE Trans.
Inform. Theory, vol. 47, pp. 520–548, Feb. 2001.
[3] G. D. Forney, Jr., “MacWilliams identities for codes on graphs,” Proc.
IEEE Inform. Theory Workshop (Taormina, Sicily), pp. 120–124, Oct.
2009. ArXiv: 0903.4207 [cs.IT].
[4] G. D. Forney, Jr., “Codes on graphs: MacWilliams identities,” submitted
to IEEE Trans. Inform. Theory, 2009. ArXiv: 0911.5508 [cs.IT].
[5] G. D. Forney, Jr., F. R. Kschischang, B. Marcus and S. Tuncel, “Iterative
decoding of tail-biting trellises and connections with symbolic dynamics,”
in Codes, Systems and Graphical Models (B. Marcus and J. Rosenthal,
eds.), pp. 239–264. New York: Springer, 2001.
[6] H. Gluesing-Luerssen and G. Schneider, “On the MacWilliams identity
for convolutional codes,” IEEE Trans. Inform. Theory, vol. 54, pp. 1536–
1550, April 2008. ArXiv: cs/0603013.
[7] H. Gluesing-Luerssen and G. Schneider, “A MacWilliams identity for
convolutional codes: The general case,” IEEE Trans. Inform. Theory, vol.
55, pp. 2920–2930, July 2009. ArXiv: 0805.3484v1 [cs.IT].
[8] J. B. Shearer and R. J. McEliece, “There is no MacWilliams identity for
convolutional codes,” IEEE Trans. Inform. Theory, vol. 23, pp. 775–776,
Nov. 1977.
V. C ONCLUSION
In summary, similarly to [1], but using tail-biting terminated
codes, we have shown that there is a MacWilliams identity
between the generating functions of the weight distributions
per unit time of a linear convolutional code C and its orthogonal code C ⊥ , and that this distribution is as useful as
the free distance spectrum for estimating code performance.
These results effectively resolve the puzzle posed by Shearer
and McEliece [8].
ACKNOWLEDGMENT
For an advance copy of [1], I am grateful to R. Johannesson.
1 Another way of reaching the same conclusion is to observe that C and
1
C2 are equivalent under a simple time-invariant, finite-memory permutation.
Therefore C1⊥ and C2⊥ are actually equivalent under the same permutation,
and thus must have precisely the same performance.
1109
Download